Hugging Face's logo

Datasets:
kesava89
/
adaptive-curriculum-tool-calling-pool

Tasks:
Text Generation
Modalities:
Tabular
Text
Languages:
English
Size:
1M - 10M
Tags:
tool-calling
function-calling
parquet
curriculum-learning
relevance-control
partition-manifest
License:
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Adaptive Curriculum Tool-Calling Pool — v2.1

A gated snapshot of the tool-calling training-data pool produced by the Adaptive Curriculum for Tool Calling sub-experiment. This is the additive v2.1 revision: it keeps the entire v1 + v2 payload and adds the six per-campaign partition manifests under manifests/partitions/. Nothing from v1 or v2 was re-encoded, recompressed, moved, or rewritten.

Access is gated. The repository uses manual gating. You must be granted access by the owner and use an approved Hugging Face token to download it.

What is new in v2.1

Six per-campaign partition manifests are added at manifests/partitions/oss_<config>_all_structs_plus_irrelevance_partition.json (one per campaign shard, matching the shard of the same config):

  1. Stratified split assignment. Each manifest is the frozen train / controller / development / audit assignment (target fractions 0.80 / 0.10 / 0.06 / 0.04) for every conversation group in that campaign shard, built with seed=20260912, stratified=true and min_per_family=5.
  2. relevance family coverage in the controller split. Every manifest reserves a non-empty controller slice for the relevance family (3,571 relevance groups per manifest), so the fixed controller probe covers the relevance control group.
  3. Reproducibility requirement. A campaign run MUST pass the matching --partition-manifest (for example manifests/partitions/oss_gemma4_E2B_all_structs_plus_irrelevance_partition.json for the oss_gemma4_E2B campaign shard). If --partition-manifest is omitted, the trainer silently falls back to its own default (cache/partition_<partition_seed>.json), which is a different, smoke-derived manifest — not the campaign manifest shipped here — so the run is no longer reproducible from this Hub snapshot.

What v2 added

  1. Relevance control group — data/irrelevance/kimi/singleturn_singlecall.parquet (35,720 rows, 277,949,252 bytes). The new relevance config exposes it.
  2. Campaign shards — campaign/oss_<config>_all_structs_plus_irrelevance.parquet (6 files, one per generator config, 1,491,773 rows total, 4,903,840,591 bytes). Each shard is that config's four oss struct cells plus the relevance cell, so a campaign row set is runnable from a single file. The new campaign config exposes all six.
  3. v2 manifest — manifests/pool_manifest_v2.json records lineage (the exact v1 revision + tag), the pinned irrelevance corpus, the label rule, the owner's human-validation attestation, and per-file sizes/sha256.

The 72 v1 Parquet cells are carried forward byte-identical (hard links); their sha256 values are unchanged from v1. v1 remains independently verifiable at its own revision/tag:

  • tag v1.0.0 → commit 0084546bdfcaa0a976eef411d6a701c37a307113
  • v1 root digest 8df27f9d44c818aecddb361e81229d617c3e213526a101bae6f7fc3b95332094

v2 also remains untouched and independently verifiable at its own revision/tag:

  • tag v2.0.0 → commit b0f869aa4888c1df4610ef406e6528cbc6028d55
  • v2 root digest 89890012e73fb69aa01047cc08aa4159477c44626bc19d0a42baa81f1e4a4b93

Vocabulary

The structure labels used throughout this dataset follow these definitions exactly:

  • conversation = a full related user conversation.
  • single turn = one user-assistant interaction, with multiple tool-calls possible.
  • multi-turn = multiple user interactions.
  • A single call explodes the data and waits for a tool response. A multi-call can or will not wait for a tool-response.

Layout 

data/<source>/<config>/<struct>.parquet         # 72 frozen v1 cells (byte-identical)
data/irrelevance/kimi/singleturn_singlecall.parquet   # v2 relevance control cell
campaign/oss_<config>_all_structs_plus_irrelevance.parquet  # v2 campaign shards (6)
manifests/
  pool_manifest.json          # per-cell v1 build record (rows, bytes, dedup)
  effective_supply.json       # cross-cell pair_id dedup accounting
  annotation_manifest.json    # annotation provenance + per-source HF revisions
  partition_20260912.json     # pool-wide stratified partition (default fallback)
  partition_smoke100.json     # small smoke partition (identical bytes to above)
  pool_manifest_v2.json       # v2 lineage, irrelevance pin, attestation, hashes
  partitions/                 # v2.1 per-campaign partition manifests (6)
    oss_<config>_all_structs_plus_irrelevance_partition.json
checksums/
  content-manifest.json       # {path, size, sha256} for every payload file
  SHA256SUMS                  # sha256␠␠path
  dataset-digest.txt          # root digest over the canonical record sequence
README.md
DATASET_CARD.md
VERSION
verify_dataset.py             # standalone verifier (stdlib + pyarrow)

Frozen v1 cross product (unchanged)

  • sources (3): oss, kimi, qwen
  • generator configs (6): gemma4_E2B, granite4_1b, granite4_350m, granite4_3b, qwen3_0.6B, qwen3_1.7B
  • structs (4): singleturn_singlecall, singleturn_multicall, multiturn_singlecall, multiturn_multicall
  • 3 × 6 × 4 = 72 cells, 3,094,219 physical rows.

Configs and splits

There are six named configs, each declaring a single train split (there is no train/test split inside the Parquet files):

config resolves
singleturn_singlecall 18 frozen v1 cells (oss + kimi + qwen)
singleturn_multicall 18 frozen v1 cells
multiturn_singlecall 18 frozen v1 cells
multiturn_multicall 18 frozen v1 cells
relevance the v2 relevance control cell
campaign the six v2 campaign shards

The four struct configs are restricted to the three v1 sources so that they keep resolving exactly the v1 cells. In v1 these were declared as the single pattern data/*/*/<struct>.parquet; that pattern would now also sweep in the new data/irrelevance/ control cell, so the struct configs use the equivalent per-source patterns data/{oss,kimi,qwen}/*/<struct>.parquet. Every v1 path is still resolved, and the relevance control group stays a separate config.

The dataset-level split assignment lives outside the pool bytes. The v1 manifests/partition_*.json files are pool-wide partitions; v2.1 adds the authoritative per-campaign manifests under manifests/partitions/. Treat the declared train split as "all physical rows", then join against the matching campaign partition manifest to obtain the actual experiment partitions.

Partition manifests (v2.1)

manifests/partitions/ holds six manifests, one per campaign shard:

manifests/partitions/oss_gemma4_E2B_all_structs_plus_irrelevance_partition.json
manifests/partitions/oss_granite4_1b_all_structs_plus_irrelevance_partition.json
manifests/partitions/oss_granite4_350m_all_structs_plus_irrelevance_partition.json
manifests/partitions/oss_granite4_3b_all_structs_plus_irrelevance_partition.json
manifests/partitions/oss_qwen3_0.6B_all_structs_plus_irrelevance_partition.json
manifests/partitions/oss_qwen3_1.7B_all_structs_plus_irrelevance_partition.json

Each manifest is the frozen stratified assignment for every conversation group in the matching campaign/oss_<config>_all_structs_plus_irrelevance.parquet shard:

Field Value
seed 20260912 (campaign partition seed, not the training seed)
stratified true
min_per_family 5 (every family keeps at least 5 groups per split)
fractions train 0.80 / controller 0.10 / development 0.06 / audit 0.04
families count, name, other, parameter, relevance, type, unlabeled, value
assignment `{conversation group id -> train

Every manifest reserves a non-empty controller slice for the relevance family (3,571 relevance groups per manifest), so the fixed controller probe covers the relevance control group. Per-manifest sizes:

Campaign manifest Rows in shard Groups train controller development audit
oss_gemma4_E2B 251,477 87,761 70,412 8,810 5,148 3,391
oss_granite4_1b 223,238 82,000 65,815 8,167 4,817 3,201
oss_granite4_350m 243,445 86,913 69,701 8,645 5,156 3,411
oss_granite4_3b 209,973 80,189 64,369 8,018 4,687 3,115
oss_qwen3_0.6B 307,826 103,534 83,065 10,293 6,117 4,059
oss_qwen3_1.7B 255,814 93,679 75,136 9,352 5,511 3,680
Total 1,491,773 534,076 428,498 53,285 31,436 20,857

A run MUST pass the matching --partition-manifest. For example, a gemma4_E2B campaign must pass manifests/partitions/oss_gemma4_E2B_all_structs_plus_irrelevance_partition.json. If --partition-manifest is omitted, the trainer silently falls back to its own default (cache/partition_<partition_seed>.json), which is a different, smoke-derived manifest rather than the campaign manifest shipped here — so the run is no longer reproducible from this snapshot and the relevance controller slice is not guaranteed. The manifests are also the mechanism that keeps a campaign's arms (seeds 42/43/44) on a shared, comparable split.

Relevance control group (v2)

The relevance control group is a single cell built from the pinned irrelevance corpus:

Field Value
Rows 35,720
Path data/irrelevance/kimi/singleturn_singlecall.parquet
Source / config / struct irrelevance / kimi / singleturn_singlecall
Upstream corpus tum-ai/toucan-irrelevance-kimi @ 07c24ef97d57e46b7232bdf7c0b71aa22bad99bd
Upstream source dataset Agent-Ark/Toucan-1.5M, subset single-turn-original
Control group relevance
Annotation band NA
Difficulty not_applicable
Chosen / rejected behaviour chosen makes zero tool calls; rejected makes one unwarranted call
Chosen / rejected equivalent False (all rows)

The band is NA because relevance rows have difficulty=not_applicable; they are a relevance control, not a difficulty-graded sample.

Label rule (verbatim from manifests/pool_manifest_v2.json):

chosen carries no error; rejected carries irrelevance_error:unexpected_tool_call:<scenario>; control_group=relevance, annotation_band=NA

Scenario mix (rejected irrelevance_error:unexpected_tool_call:<scenario>):

Scenario Rows
unavailable_capability 16,431
missing_user_argument 10,992
insufficient_tool_schema 8,297
Total 35,720

Human-validation attestation

manifests/pool_manifest_v2.json records the owner's attestation that human validation of the irrelevance rows is complete. Reproduced exactly, including its caveat:

{
  "human_validation": {
    "attested_by": "dataset owner",
    "attested_on": "2026-09-14",
    "caveat": "the corpus revision is unchanged and its own report.json still reads 'awaiting_human_and_downstream_validation'; this is the owner's attestation, recorded here for provenance",
    "status": "attested_complete"
  }
}

The caveat is important: the upstream corpus revision is unchanged and its own report.json still reads awaiting_human_and_downstream_validation. This is the dataset owner's attestation, recorded for provenance — it is not an upstream validation status change.

Numbers

Quantity Value
Payload files (total) 95
Parquet files (total) 79
Non-Parquet manifests (v1 + v2) 6
Partition manifests (v2.1) 6
Release metadata 4 (README.md, DATASET_CARD.md, VERSION, verify_dataset.py)
Frozen v1 cells 72 (3,094,219 rows, 66,249,143,184 bytes)
Relevance cells 1 (35,720 rows, 277,949,252 bytes)
Campaign shards 6 (1,491,773 rows, 4,903,840,591 bytes)
Physical rows, all Parquet 4,621,712
Partition groups assigned (all 6 manifests) 534,076
Relevance groups in controller (per manifest) 3,571
Effective rows after cross-cell pair_id dedup (v1 cells) 2,214,068
Generator configs 6
Sources (v1) 3
Structures 4
Tokenizer ibm-granite/granite-4.0-350m
Schema one logical 33-column schema (see caveat)

The exact byte total and root digest over the full v2.1 payload are in checksums/content-manifest.json and checksums/dataset-digest.txt.

Schema

All v1 files share one logical 33-column schema. Because the v1 pool was assembled by more than one build pass, the physical Arrow files carry four projections of it: some cells omit the optional pair_provenance column, and a few declare toolcall_mode as an all-null (null-typed) column. Both variants are part of the frozen snapshot and are verified by verify_dataset.py. Column names include conv_id, turn_id, conv_tracking_id, struct, source, config, pair_id, pair_provenance, curriculum_reward_diff, toolrl_reward_diff, the BFCL error-type/family lists, and the p_ids/c_ids/r_ids plus P/C/R packing columns.

The relevance cell has 32 of those columns (it omits pair_provenance) and carries all-null typed columns (chosen_bfcl_error_types, chosen_bfcl_families, toolrl_reward_diff); those variants are part of the frozen relevance corpus. The campaign shards are uniform: all 33 columns, with pair_provenance present.

Compression / file format

Files are the original canonical bytes. Compression varies by column: data columns are SNAPPY-compressed, while some columns are stored UNCOMPRESSED. Uniform SNAPPY should not be assumed. Do not recompress the files if you intend to match the published checksums.

Provenance pins

  • BFCL / Gorilla upstream commit: bfcl_gorilla_commit = f7cf7359b7ac615a0b294831c5ba2bc95ee4a000
  • Per-source Hugging Face revisions: recorded in manifests/annotation_manifest.json under sources_hf_revision (oss: 6dd8926cb6b029f5b3141b59b0b1ed3d4661b1f8, kimi: 0fcd6729cbc7873cc4eafee91e84cd54fe4d44e6, qwen: 3e98d3fb8f179f598f52d48b49d907e171e249c9).
  • Irrelevance corpus: tum-ai/toucan-irrelevance-kimi @ 07c24ef97d57e46b7232bdf7c0b71aa22bad99bd (manifests/pool_manifest_v2.json → irrelevance_corpus).
  • v1 lineage: commit 0084546bdfcaa0a976eef411d6a701c37a307113, tag v1.0.0 (manifests/pool_manifest_v2.json → derived_from).
  • v2 lineage: commit b0f869aa4888c1df4610ef406e6528cbc6028d55, tag v2.0.0, root digest 89890012e73fb69aa01047cc08aa4159477c44626bc19d0a42baa81f1e4a4b93 (carried forward byte-identical into v2.1).
  • Partition manifests: manifests/partitions/oss_<config>_all_structs_plus_irrelevance_partition.json (seed 20260912, stratified, min_per_family=5).
  • Annotation format version: annotation_format_version = 1 (manifests/annotation_manifest.json).
  • Build record: manifests/pool_manifest.json (v1 tokenizer, per-cell rows/bytes/dedup).

Loading

Always pin the full commit SHA of the v2.1 revision, not a branch name, so you load exactly these bytes:

from datasets import load_dataset

REVISION = "<FULL_V2_1_COMMIT_SHA>"  # 40 hex chars

ds = load_dataset(
    "kesava89/adaptive-curriculum-tool-calling-pool",
    "singleturn_singlecall",          # or any of the six configs
    revision=REVISION,
)
print(ds["train"])                    # 18 frozen v1 cells of this structure

rel = load_dataset(
    "kesava89/adaptive-curriculum-tool-calling-pool",
    "relevance",
    revision=REVISION,
)

To load the runnable campaign shards:

camp = load_dataset(
    "kesava89/adaptive-curriculum-tool-calling-pool",
    "campaign",
    revision=REVISION,
)

To make a campaign reproducible, download the matching partition manifest and pass it to every arm of the campaign:

huggingface-cli download kesava89/adaptive-curriculum-tool-calling-pool \
  --repo-type dataset --revision "$REVISION" \
  --include "manifests/partitions/oss_gemma4_E2B_all_structs_plus_irrelevance_partition.json" \
  --local-dir ./snapshot
# then, for every arm of the campaign:
#   --partition-manifest ./snapshot/manifests/partitions/oss_gemma4_E2B_all_structs_plus_irrelevance_partition.json

Because the repo is gated, the token you pass to load_dataset / huggingface-cli must belong to an account that has been granted access. The owner account is always allowed.

To fetch only specific cells, use data_files with load_dataset("parquet", ...) or download individual paths directly.

Integrity verification

The published payload is fully self-describing. After downloading, run the verifier shipped in the repository:

python verify_dataset.py --root /path/to/downloaded/snapshot

It auto-detects the release from VERSION and checks:

  • the exact 72 frozen v1 cells, plus the relevance cell and the six campaign shards (79 Parquet files);
  • every file's size and SHA-256 against checksums/content-manifest.json, and the root digest in checksums/dataset-digest.txt;
  • checksums/SHA256SUMS consistency;
  • exactly 3,094,219 frozen v1 physical rows, 35,720 relevance rows, and 1,491,773 campaign rows (metadata-only);
  • one logical 33-column schema across the frozen v1 cells, with the known physical projections; the relevance cell's canonical-minus-pair_provenance schema; and the canonical 33-column campaign-shard schema;
  • for v2.1, the six manifests/partitions/ campaign manifests: present and well formed, seed=20260912, stratified, min_per_family=5, the four-way train/controller/development/audit split, split counts that agree with the assignment, and a non-empty relevance family in the controller split.

The root digest for this snapshot and the final commit SHA are recorded in the operational runbook tools/HF_PUBLISH_README.md in the parent project, and can be compared against checksums/dataset-digest.txt.

Older revisions stay verifiable. Verify v1 or v2 by pinning their own revision or tag; each snapshot ships the verify_dataset.py of its era and checks its own checksums:

huggingface-cli download kesava89/adaptive-curriculum-tool-calling-pool \
  --repo-type dataset --revision v1.0.0 --local-dir ./snapshot-v1
python ./snapshot-v1/verify_dataset.py --root ./snapshot-v1

huggingface-cli download kesava89/adaptive-curriculum-tool-calling-pool \
  --repo-type dataset --revision v2.0.0 --local-dir ./snapshot-v2
python ./snapshot-v2/verify_dataset.py --root ./snapshot-v2

License and redistribution

This is a research snapshot, not a blanket license grant. license: other.

  • The BFCL / Gorilla upstream data that seeds this pool is Apache-2.0 (pinned at commit f7cf7359b7ac615a0b294831c5ba2bc95ee4a000). Apache-2.0 requires preserving attribution and license notices for the upstream-derived portions.
  • The relevance control group is derived from tum-ai/toucan-irrelevance-kimi (revision pinned above), sourced from Agent-Ark/Toucan-1.5M (single-turn-original). Redistribution rights depend on the terms of that dataset and its upstream sources.
  • The remaining content is model-generated (from the oss, kimi, and qwen source models listed under Provenance pins) and was annotated with BFCL reward signals. Redistribution rights for model-generated outputs depend on the terms of each generating model and the upstream datasets.

TODO (owner action, required before granting access): The dataset owner must confirm the license and redistribution rights for the model-generated source data, verify the terms of each source model/dataset revision listed in manifests/annotation_manifest.json and of the irrelevance corpus pinned in manifests/pool_manifest_v2.json, add the appropriate upstream NOTICE/ attribution files, and replace this section with a definitive license statement. Until then this dataset must remain gated to approved users only.

Citation

If you use this pool, cite the parent Adaptive Curriculum for Tool Calling sub-experiment and the BFCL / Gorilla benchmark (Apache-2.0).
Downloads last month
44
Total file size:
71.5 GB